Sliding ferroelectricity (FE) in two-dimensional (2D) materials offers a promising route to ultrathin, low-power, nonvolatile devices, yet in pristine homobilayers it is typically weak and confined to a small subset of known 2D materials. This limitation prompts a central question: Can sliding FE be reliably induced and substantially enhanced in otherwise nonferroelectric homobilayers through a controllable, experimentally feasible intervention? Here, we combine high-throughput first-principles calculations with machine learning to establish atomic intercalation as a versatile strategy for activating and amplifying sliding FE. Notably, we identify pristine InP and GaP as record performers, exhibiting out-of-plane polarizations up to 21.0 pC/m, approximately ten times greater than experimentally synthesized BN and over 45 times higher than $\mathrm{Mo}{\mathrm{S}}_2$, while maintaining moderate switching barriers of just 13.6 and 32.0 meV/f.u. Intercalation with Cu, Ag, and Au expands the pool from 106 pristine sliding FE systems to 591, including 212 activated from initially nonferroelectric states. To accelerate exploration of this enlarged configurational space, we develop a descriptor-free graph neural network incorporating message passing and global attention, enabling direct polarization prediction from atomic configurations. This integrated computational–machine learning framework delivers a robust and tunable pathway for designing high-performance sliding ferroelectrics and a blueprint for targeted discovery of functional 2D materials.

Close

PrevNext

View figure in article

• Ferroelectricity
• First-principles calculations
• Intercalation
• 2-dimensional systems
• Density functional theory
• High-throughput calculations
• Machine learning